Script to reproduce years based on a model trained with random points¶
Importing¶
In [ ]:
import xarray as xr
import numpy as np
import pandas as pd
import matplotlib.pyplot as plt
from sklearn.pipeline import make_pipeline
from sklearn.preprocessing import StandardScaler
from sklearn.preprocessing import QuantileTransformer
from sklearn.compose import TransformedTargetRegressor
from sklearn.neural_network import MLPRegressor
from sklearn.ensemble import AdaBoostRegressor
from sklearn.ensemble import BaggingRegressor
from sklearn.metrics import root_mean_squared_error as rmse
from tqdm import tqdm
import dill
import random
import salishsea_tools.viz_tools as sa_vi
Datasets Preparation¶
In [ ]:
def datasets_preparation(dataset, dataset2):
drivers = np.stack([np.ravel(dataset['Temperature_(0m-15m)']),
np.ravel(dataset['Temperature_(15m-100m)']),
np.ravel(dataset['Salinity_(0m-15m)']),
np.ravel(dataset['Salinity_(15m-100m)']),
np.ravel(dataset2['Summation_of_solar_radiation']),
np.ravel(dataset2['Mean_wind_speed']),
np.ravel(dataset2['Mean_air_temperature'])
])
indx = np.where(~np.isnan(drivers).any(axis=0))
drivers = drivers[:,indx[0]]
diat = np.ravel(dataset['Diatom'])
diat = diat[indx[0]]
return(drivers, diat, indx)
Regressor¶
In [ ]:
def regressor (inputs, targets):
inputs = inputs.transpose()
model = TransformedTargetRegressor(regressor=make_pipeline(StandardScaler(),MLPRegressor(alpha=0.001, learning_rate='invscaling', epsilon=1e-07, hidden_layer_sizes=[100,100])),
transformer=QuantileTransformer())
regr = BaggingRegressor(model, n_estimators=10, n_jobs=10).fit(inputs,targets)
return(regr)
Regressor 2¶
In [ ]:
def regressor2 (inputs, targets, variable_name):
inputs2 = inputs.transpose()
outputs_test = regr.predict(inputs2)
m = scatter_plot(targets, outputs_test, variable_name)
r = np.round(np.corrcoef(targets, outputs_test)[0][1],3)
rms = rmse(targets, outputs_test)
return (r, rms, m)
Regressor 3¶
In [ ]:
def regressor3 (inputs, targets):
inputs2 = inputs.transpose()
outputs_test = regr.predict(inputs2)
# compute slope m and intercept b
m, b = np.polyfit(targets, outputs_test, deg=1)
r = np.round(np.corrcoef(targets, outputs_test)[0][1],3)
rms = rmse(targets, outputs_test)
return (r, rms, m)
Regressor 4¶
In [ ]:
def regressor4 (inputs, targets, variable_name):
inputs2 = inputs.transpose()
outputs = regr.predict(inputs2)
# Post processing
indx2 = np.full((len(diat_i.y)*len(diat_i.x)),np.nan)
indx2[indx[0]] = outputs
model = np.reshape(indx2,(len(diat_i.y),len(diat_i.x)))
m = scatter_plot(targets, outputs, variable_name + str(dates[i].date()))
# Preparation of the dataarray
model = xr.DataArray(model,
coords = {'y': diat_i.y, 'x': diat_i.x},
dims = ['y','x'],
attrs=dict( long_name = variable_name + "Concentration",
units="mmol m-2"),)
plotting3(targets, model, diat_i, variable_name)
Printing¶
In [ ]:
def printing (targets, outputs, m):
print ('The amount of data points is', outputs.size)
print ('The slope of the best fitting line is ', np.round(m,3))
print ('The correlation coefficient is:', np.round(np.corrcoef(targets, outputs)[0][1],3))
print (' The root mean square error is:', rmse(targets,outputs))
Scatter Plot¶
In [ ]:
def scatter_plot(targets, outputs, variable_name):
# compute slope m and intercept b
m, b = np.polyfit(targets, outputs, deg=1)
printing(targets, outputs, m)
fig, ax = plt.subplots(2, figsize=(5,10), layout='constrained')
ax[0].scatter(targets,outputs, alpha = 0.2, s = 10)
lims = [np.min([ax[0].get_xlim(), ax[0].get_ylim()]),
np.max([ax[0].get_xlim(), ax[0].get_ylim()])]
# plot fitted y = m*x + b
ax[0].axline(xy1=(0, b), slope=m, color='r')
ax[0].set_xlabel('targets')
ax[0].set_ylabel('outputs')
ax[0].set_xlim(lims)
ax[0].set_ylim(lims)
ax[0].set_aspect('equal')
ax[0].plot(lims, lims,linestyle = '--',color = 'k')
h = ax[1].hist2d(targets,outputs, bins=100, cmap='jet',
range=[lims,lims], cmin=0.1, norm='log')
ax[1].plot(lims, lims,linestyle = '--',color = 'k')
# plot fitted y = m*x + b
ax[1].axline(xy1=(0, b), slope=m, color='r')
ax[1].set_xlabel('targets')
ax[1].set_ylabel('outputs')
ax[1].set_aspect('equal')
fig.colorbar(h[3],ax=ax[1], location='bottom')
fig.suptitle(variable_name)
plt.show()
return (m)
Plotting¶
In [ ]:
def plotting(variable, name):
plt.plot(years,variable, marker = '.', linestyle = '')
plt.xlabel('Years')
plt.ylabel(name)
plt.show()
Plotting 2¶
In [ ]:
def plotting2(variable,title):
fig, ax = plt.subplots()
scatter= ax.scatter(dates,variable, marker='.', c=pd.DatetimeIndex(dates).month)
ax.legend(handles=scatter.legend_elements()[0], labels=['February','March','April'])
fig.suptitle('Daily ' + title + ' (15 Feb - 30 Apr)')
fig.show()
Plotting 3¶
In [ ]:
def plotting3(targets, model, variable, variable_name):
fig, ax = plt.subplots(2,2, figsize = (10,15))
cmap = plt.get_cmap('cubehelix')
cmap.set_bad('gray')
variable.plot(ax=ax[0,0], cmap=cmap, vmin = targets.min(), vmax =targets.max(), cbar_kwargs={'label': variable_name + ' Concentration [mmol m-2]'})
model.plot(ax=ax[0,1], cmap=cmap, vmin = targets.min(), vmax = targets.max(), cbar_kwargs={'label': variable_name + ' Concentration [mmol m-2]'})
((variable-model) / variable * 100).plot(ax=ax[1,0], cmap=cmap, cbar_kwargs={'label': variable_name + ' Concentration [percentage]'})
plt.subplots_adjust(left=0.1,
bottom=0.1,
right=0.95,
top=0.95,
wspace=0.35,
hspace=0.35)
sa_vi.set_aspect(ax[0,0])
sa_vi.set_aspect(ax[0,1])
sa_vi.set_aspect(ax[1,0])
ax[0,0].title.set_text(variable_name + ' (targets)')
ax[0,1].title.set_text(variable_name + ' (outputs)')
ax[1,0].title.set_text('targets - outputs')
ax[1,1].axis('off')
fig.suptitle(str(dates[i].date()))
plt.show()
Training (Random Points)¶
In [ ]:
ds = xr.open_dataset('/data/ibougoudis/MOAD/files/integrated_model_var_old.nc')
ds2 = xr.open_dataset('/data/ibougoudis/MOAD/files/external_inputs.nc')
# ds = ds.isel(time_counter = (np.arange(0, len(ds.time_counter),2)),
# y=(np.arange(ds.y[0], ds.y[-1], 5)),
# x=(np.arange(ds.x[0], ds.x[-1], 5)))
# ds2 = ds2.isel(time_counter = (np.arange(0, len(ds2.time_counter),2)),
# y=(np.arange(ds2.y[0], ds2.y[-1], 5)),
# x=(np.arange(ds2.x[0], ds2.x[-1], 5)))
dataset = ds.sel(time_counter = slice('2007', '2020'))
dataset2 = ds2.sel(time_counter = slice('2007', '2020'))
drivers, diat, _ = datasets_preparation(dataset, dataset2)
regr = regressor(drivers, diat)
Daily Maps¶
In [ ]:
ds = ds.sel(time_counter = slice('2021', '2023'))
ds2 = ds2.sel(time_counter = slice('2021', '2023'))
dates = pd.DatetimeIndex(ds['time_counter'].values)
maps = random.sample(range(0,len(ds.time_counter)),10)
for i in tqdm(maps):
dataset = ds.isel(time_counter=i)
dataset2 = ds2.isel(time_counter=i)
drivers, diat, indx = datasets_preparation(dataset, dataset2)
diat_i = dataset['Diatom']
regressor4(drivers, diat, 'Diatom ')
0%| | 0/10 [00:00<?, ?it/s]
The amount of data points is 46479 The slope of the best fitting line is 0.279 The correlation coefficient is: 0.43 The root mean square error is: 0.20322512036762408
10%|█ | 1/10 [00:15<02:22, 15.84s/it]
The amount of data points is 46479 The slope of the best fitting line is 0.525 The correlation coefficient is: 0.693 The root mean square error is: 0.05155939498026469
20%|██ | 2/10 [00:29<01:57, 14.71s/it]
The amount of data points is 46479 The slope of the best fitting line is 0.079 The correlation coefficient is: 0.051 The root mean square error is: 0.2110631584483521
30%|███ | 3/10 [00:42<01:37, 13.88s/it]
The amount of data points is 46479 The slope of the best fitting line is 0.477 The correlation coefficient is: 0.773 The root mean square error is: 0.06722725285034359
40%|████ | 4/10 [00:54<01:17, 12.98s/it]
The amount of data points is 46479 The slope of the best fitting line is 0.191 The correlation coefficient is: 0.218 The root mean square error is: 0.15083785204669917
50%|█████ | 5/10 [01:05<01:02, 12.44s/it]
The amount of data points is 46479 The slope of the best fitting line is 0.549 The correlation coefficient is: 0.641 The root mean square error is: 0.0585947408214611
60%|██████ | 6/10 [01:17<00:49, 12.26s/it]
The amount of data points is 46479 The slope of the best fitting line is 0.506 The correlation coefficient is: 0.697 The root mean square error is: 0.058046000453268874
70%|███████ | 7/10 [01:29<00:36, 12.01s/it]
The amount of data points is 46479 The slope of the best fitting line is 0.314 The correlation coefficient is: 0.589 The root mean square error is: 0.17371503944982114
80%|████████ | 8/10 [01:40<00:23, 11.79s/it]
The amount of data points is 46479 The slope of the best fitting line is 0.278 The correlation coefficient is: 0.528 The root mean square error is: 0.10248697826391519
90%|█████████ | 9/10 [01:51<00:11, 11.70s/it]
The amount of data points is 46479 The slope of the best fitting line is 0.853 The correlation coefficient is: 0.848 The root mean square error is: 0.07959479131658387
100%|██████████| 10/10 [02:03<00:00, 12.35s/it] 100%|██████████| 10/10 [02:03<00:00, 12.35s/it]
Other Years (Anually)¶
In [ ]:
years = range (2021,2024)
for year in tqdm(range (2021,2024)):
dataset = ds.sel(time_counter=str(year))
dataset2 = ds2.sel(time_counter=str(year))
drivers, diat, _ = datasets_preparation(dataset, dataset2)
r, rms, m = regressor2(drivers, diat, 'Diatom ' + str(year))
0%| | 0/3 [00:00<?, ?it/s]/tmp/ipykernel_14146/1980467486.py:4: RankWarning: Polyfit may be poorly conditioned m, b = np.polyfit(targets, outputs, deg=1)
The amount of data points is 3485925 The slope of the best fitting line is 0.447 The correlation coefficient is: 0.645 The root mean square error is: 0.1378248039117271
33%|███▎ | 1/3 [01:19<02:39, 79.64s/it]/tmp/ipykernel_14146/1980467486.py:4: RankWarning: Polyfit may be poorly conditioned m, b = np.polyfit(targets, outputs, deg=1)
The amount of data points is 3485925 The slope of the best fitting line is 0.47 The correlation coefficient is: 0.557 The root mean square error is: 0.12422049887596191
67%|██████▋ | 2/3 [02:27<01:12, 72.79s/it]/tmp/ipykernel_14146/1980467486.py:4: RankWarning: Polyfit may be poorly conditioned m, b = np.polyfit(targets, outputs, deg=1)
The amount of data points is 3485925 The slope of the best fitting line is 0.5 The correlation coefficient is: 0.371 The root mean square error is: 0.16468230161096423
100%|██████████| 3/3 [03:36<00:00, 72.16s/it] 100%|██████████| 3/3 [03:36<00:00, 72.16s/it]
Other Years (Daily)¶
In [ ]:
r_all2 = np.array([])
rms_all2 = np.array([])
slope_all2 = np.array([])
for i in tqdm(range (0, len(ds.time_counter))):
dataset = ds.isel(time_counter=i)
dataset2 = ds2.isel(time_counter=i)
drivers, diat, _ = datasets_preparation(dataset, dataset2)
r, rms, m = regressor3(drivers, diat)
r_all2 = np.append(r_all2,r)
rms_all2 = np.append(rms_all2,rms)
slope_all2 = np.append(slope_all2,m)
plotting2(r_all2, 'Correlation Coefficients')
plotting2(rms_all2, 'Root Mean Square Errors')
plotting2(slope_all2, 'Slope of the best fitting line')
100%|██████████| 225/225 [35:02<00:00, 9.34s/it]
In [ ]: